Replace the doc-style checker with a documentation maintenance skill

[MirzaMerdovic]

Summary

The doc-style skill now maintains documentation. It no longer checks prose against fixed
rules. When Claude changes code, the skill directs it to find the comments, docstrings, and
documents that describe the changed code. Claude edits, consolidates, or deletes that prose
before it adds more. This change removes the Bash checker and its tests.

Changes

• skills/doc-style/SKILL.md holds instructions only:
  • a six-step workflow scaled to the size of the change;
  • the edit order: delete, correct, consolidate, replace, add;
  • current state versus change history;
  • rules for comments and docstrings;
  • the writing style, with no sentence limit and no word list.
• skills/doc-style/examples/maintenance.md gives eight maintenance decisions, with the
  text before and after.
• skills/doc-style/examples/rewrites.md gives sentence rewrites, each judged in context.
• hooks/hooks.json adds a SessionStart hook that prints snippets/CLAUDE.md into the
  context of each session. In eval runs on Sonnet without the hook, Claude loaded the skill
  in 0 of 7 tasks that asked only for a code change. With the hook, it loaded the skill in
  every run.
• evals/ holds nine cases for claude plugin eval. Each case copies a small workspace
  with scaffold.sh and grades the files that Claude leaves.
• The change removes these files and directories:
  • skills/doc-style/scripts/check, rules.md, and deviations.md;
  • .doc-style and commands/doc-style-check.md;
  • tests/ and .github/workflows/mutate.yml;
  • mise.toml and mise.lock.
• CI runs only claude plugin validate. mise.ci.toml still pins the Claude Code version
  for that step.

Compatibility

• The /haddock:doc-style-check command is removed. A .doc-style file has no effect.
• The plugin version stays 0.4.0. An installed copy can keep the old content until the
  version changes.
• The hook adds about 190 tokens to every session that has the plugin enabled.

Validation

• claude plugin validate . passes with Claude Code 2.1.292, the version that CI pins.

• shellcheck and shfmt -d pass on the scaffold.sh files.

• claude plugin eval . --scaffold --allow-tools Edit Write ran each case with and
  without the plugin. A score is the fraction of graders that passed:

  Model	Runs per case and arm	Score with plugin	Score without plugin	Skill loaded
  Opus	2	0.99	0.76	18 of 18
  Sonnet	1	0.94	0.79	9 of 9

  Without the plugin, Claude did not do these updates:

  • It did not correct a README paragraph that a behavior change made wrong.
  • It did not remove change history from a code comment.
  • It did not update a README table that the change affected.
  • It did not consolidate two documents that stated the same fact.

• In a three-turn claude -p --resume session on Sonnet, Claude loaded the skill once. It
  did not load the skill again for the two later code changes.

Known limitations

• Sonnet did not consolidate the duplicated documents in any of 9 runs. Opus consolidated
  them in 6 of 6 runs.
• Sonnet updated a README table that names no changed identifier in 2 of 6 runs.
• The eval cases do not grant Bash, so the evals do not test the instruction to read the
  commit history.
• The evals do not run in CI. Each run starts Claude sessions with the credentials of the
  person who runs it.
• Most results come from 1 or 2 runs per case. The LLM graders use Haiku.
• No test covers the hook after compaction.